Experiments

In order to prepare the autonomous operation of the full mission several experiments have been performed. The corresponding experiments are listed in Table 6.1. The scope of the experiments reflects an incremental increase of complexity to finally achieve the full baseline mission.

Uncooperative Reconfiguration

Attaching SherpaTT’s manipulator to a payload and placing a payload onto another maleEMI, e.g., by attaching it to Coyote III, are essential elements to demonstrate the capability for unco-operative reconfiguration. The trials are listed in TableI.1in AppendixI. All 32 trials involve payload pickup with SherpaTT’s manipulator and a subset of 12 experiments also involves the placing of a payload onto Coyote III. Both, pickup and place, rely on the same docking proce-dure although a different parametrisation for the camera calibration has to be used - depending on theEMIcamera which is used for the visual servoing approach. Camera mounting, camera calibration, marker placement, marker pose detection, and the manipulator itself influence the general accuracy of the approach. To reduce the number and impact of external error sources, (intermediate) target poses are defined as relative poses with respect to the identified marker poses. These target poses are identified through the following procedure:

• attach end effector to topEMIof payload and position payload on ground

Table 6.1:Experiments with corresponding descriptions.

Experiment Description

D-0X Pick up of a sampling module, human operator triggers the procedures D-1X Pick up of a sampling module and placing it back onto Coyote III, human

operator trigger the procedures and defines waypoints

RS-VXX Sequence of payload pick and placing it back onto Coyote III, involving Coy-ote III to autonomously approach SherpaTT from a known location and re-turning to this location, SherpaTT triggers all action of a predefined proce-dure

FM-IXX Integration testing of the performance of the fully autonomous execution of the baseline mission, SherpaTT triggers all actions of a predefined procedure FM-VXX Testing of the fully autonomous execution of the baseline mission, SherpaTT

triggers all actions of a predefined procedure

• detach end effector and move end effector in z-direction upwards for predefined distance, the detected marker position is denoted byinner alignment pose, and the moved distance the corresponding distance to command linear movement for blind docking from this pose

• move end effector in z-direction upwards for a predefined distance, the detection marker pose is denoted by outer alignment poseand the moved distance is again registered for blind linear movement

This teaching process results in two target poses, one for the outer and one for the inner align-ment.

Pick and place are initiated by a search procedure of the manipulator. Although the posi-tion of the target might not be exactly known, the target is expected to be encountered in the workspace of the manipulator at SherpaTT’s front. The search procedure first performs a spiral movement (cf. Figure6.17) and continues to align to the outer marker set once a marker has been found. If the search procedure cannot find a target marker, the reconfiguration procedure is aborted. After alignment to the outer marker board, a predefined linear movement accord-ing to the previously mentioned end effector teachaccord-ing procedure is performed. Subsequently, the alignment to the inner marker set is initiated. A successful inner alignment follows again a linear movement to attach the end effector to theEMI. Force torque values which act on the end effector are monitored and if a threshold is exceeded, the docking procedure is aborted and the manipulator returns linearly to a previous pose (defined by a minimum distance to the current pose). Figure6.18illustrates the entered task states of the component which controls the pickup of a sampling module (0 to 200 seconds) and the placing of a sampling module (500-800 seconds) for trial FM-V33. The general control approach is slow since the camera in the end effector and femaleEMIsproduces images with an approximate frequency of 1 Hz; Fig-ure6.18illustrates these low frequency transitions between marker detection and alignment.

The lift payload activity corresponds to a linear movement of the SherpaTT’s end effector in z direction. Depending upon theEMI’s locking state the end effector might either move with or without an attached payload. The locking status is controlled by another component, so that the activities have to be synchronised by a high-level action. Figure6.18shows the significantly longer duration of the placement of the payload compared to the initial placement. For trial FM-V33 the long duration of the placement is due to an outstretched manipulator as shown in Figure6.19. The wide opening angle of the manipulator’s joint 3 results in a increased position

Figure 6.17:Position and force profile during the placing of a sampling module (FM-V33).

error, since play in the first joint has a greater impact in this situation compared to using the end effector closer to the manipulator’s base. The combined performance of the pickup and place procedure is shown in Figure6.20a. The procedures have a combined success rate of 0.84 based on 20 pickup and 12 place actions. The outer alignment is the fast and reliable part of the procedure. However, outer alignment also does not require a high precision and allows for position tolerance on all axes of 2 cm and for the orientation a 8^◦tolerance for the error of roll, pitch and yaw respectively. The outer alignment is only supposed to establish the start condi-tions for the precise inner alignment, which requires the inner alignment position to be met with position accuracy of 0.8 mm and allows a maximum orientation error of 0.5^◦on each axis.

To achieve an end effector pose which remains within these tolerances, inner alignment can take up to several minutes until completion. The average duration for all types of alignment remains within a 40 s bound (cf. Figure6.20b).

FM-VXX: An Autonomous Mission with a Reconfigurable Multi-Robot System

The mission illustrated in Figure6.14is the basis for a reference experiment to illustrate the feasibility for the autonomous operation of a reconfigurable multi-robot system. The mis-sion involves SherpaTT, Coyote III and a sampling module and Coyote III in combination with a sampling payload approaches SherpaTT. SherpaTT extracts the sampling module from the composite system, and Coyote III retreats to its initial position. Coyote III moves to a target waypoint and waits for the composite agent consisting of SherpaTT and the sampling module to complete a soil sampling activity. Upon completion of this activity Coyote III is ordered back to take the sampling module and again retreat to the initial position.

Figure 6.18: Task state transitions for payload pickup and payload placement during experi-ment FM-V33.

(a) Aligning for payload pickup in the workspace between SherpaTT’s legs.

(b)During connecting to payload and lift-ing, manipulator operates with almost right angle.

(c)Aligning the payload placement in the workspace between SherpaTT’s legs.

(d)Alignment for payload placement with outstretched manipulator, leading to long alignment duration.

Figure 6.19: Exchange of a sampling module during experiment (FM-V33), pick and place from two perspectives.

(a)Inner alignment over all performed recon-figuration actions.

(b)Alignment duration of success-ful reconfiguration actions.

Figure 6.20:Alignment characteristics over all pick and place actions.

(a)Alignment to the outer marker set. (b)Alignment to the inner marker set.

Figure 6.21:Alignment characteristics for picking a payload.

(a)Alignment to the outer marker set. (b)Alignment to the inner marker set.

Figure 6.22:Alignment characteristics for placing a payload.

Table 6.2:High level action sequence for the mission in experiment FM-VXX.

# Executor Action name Action description

1 SherpaTT generate map Trigger the initial map generation on SherpaTT 2 Coyote III generate map Trigger the initial map generation on Coyote III 3 Coyote III move to target marker Coyote III visually identifies the relative pose of a

marker attached to SherpaTT and approaches this pose using a generated spline

4 SherpaTT pickup payload pickup of a payload is initiated, involving search for the top interface, attaching and extraction of the payload

5 Coyote III move blind backward Coyote III moves linearly 1.7 m straight backwards 6 Coyote III move to relative target Coyote III moves to a waypoint relative to SherpaTT,

here to relative coordinates x: 5 m, y: 2 m, trajectory planning is performed on the distributed map 7 SherpaTT pickup soil sample activation of the payload sampling sequence 8 Coyote III move to relative target Coyote III moves back to its origin pose,

com-manded as waypoint relative to SherpaTT, here rel-ative position x: 3 m y: 0 m

9 Coyote III move to target marker Coyote III identifies the relative pose of a marker at-tached to SherpaTT and approaches the pose 10 SherpaTT place payload SherpaTT places payload onto Coyote III

11 Coyote III move blind backward Coyote III moves linearly 1.7 m straight backwards

The FM-VXX experiment presents a reconfiguration sequence which can be used as a template for other multi-robot missions. The complete mission is controlled and monitored by Sher-paTT, i.e., SherpaTT acts as master while Coyote III acts as a slave and performs commands which it receives from SherpaTT. The initial relative poses of Coyote III and SherpaTT are known to the robots so that a global reference frame exists. The global reference frame is es-sential for the use of the distributed SLAM. Both robots run distributed SLAM and like for the tests described in Section6.1localised pointclouds and corresponding spatial constraints are exchanged between the robots. The distributed SLAM version used for this experiment also exchanges footprint samples, i.e., the current pose of a robot plus the information about the current footprint size of the robot. The continuous exchange of footprint data allows to mask robots in received pointclouds, so that another robot is not perceived as a permanent obstacle in a computed map. Furthermore, a footprint might change during the mission for a robot like SherpaTT.

The performed mission consists of the high level action sequence listed in Table6.2. Only the backup movement of Coyote III is defined using a predefined distance in order to guarantee that subsequent activities of SherpaTT and Coyote III can be safely executed. Otherwise, the mission approach tries to exploit the shared map and waypoints are defined as relative poses with respect to SherpaTT’s (other rather the current commanding robot’s) currently assumed pose. SherpaTT transforms these poses according to its current pose, so that Coyote III is still commanded to absolute coordinates.

In preparatory experiments FM-I1X and FM-I2X Coyote III returned to the origin and initiated a blind offset movement towards SherpaTT; this offset movement assumed a sufficiently precise pose of Coyote III at this stage. The experiments showed, however, that Coyote III did return with an orientation error to the origin. To allow for a more reliable approach to SherpaTT,

Figure 6.23:Coyote III camera view which allows to detects the position of SherpaTT with the help of an artificial marker (FM-V33).

Coyote III’s moves towards a target marker attached to SherpaTT in FM-VXX. If Coyote III is not able to detect the marker it moves incrementally forward - 0.25 m in straight line - assum-ing that this change of location improves the likelihood of detectassum-ing the marker. The approach is aborted after 4 iterations. Figure 6.23 illustrates Coyote III’s successful detection of the marker attached to SherpaTT during trial FM-V33. The design of the approach assumes that the marker will be visible to Coyote III either directly after returning to the location requested by SherpaTT or after moving Coyote III incrementally forward. The final series of experiments showed that these assumptions did not hold, and that a robust approach requires an additional search behaviour for Coyote III to guarantee for a successful cooperative docking procedure.

Robot-to-Robot Communication Robot-to-robot communication is required for command-ing robots and for exchangcommand-ing data as part of the distributed mappcommand-ing approach. SherpaTT acts as master and sends action commands to Coyote III. Since the activities of both systems have to be synchronised, the implemented communication protocol allows a robot to com-municate status changes of each action back to the robot requesting an action. The effect is illustrated in Figure6.24by the robot specific action sample count. Coyote III shows a much higher action message volume, as a result of communicating the action status, while SherpaTT only sends action commands.

The exchange of footprint data between both robots leads to a continuous data stream between both systems. It can also be seen that the footprint sample count per message exchange lies at 25 samples and could be reduced to save bandwidth. Localised pointclouds are only exchanged when a system is active and Figure6.24illustrates the messaging activity, where the transfer of localised pointclouds corresponds to messages of multiple MB.

The general communication volume is visualised according to the used communication chan-nel. The communication channel for action messages is named after the robot, while additional channels for direct subsystem communication come with an additional suffix. The suffix based naming schema simplifies the design of multicast groups for the distributed communication architecture. Sending a message to ’.*-sensors’ directs the message to all receivers where the name matches this pattern. The overall communication data volume is approximately 47 MB for data sent from SherpaTT and 73 MB for data sent from Coyote III, leading to a data rate of 273 MB/h for Coyote III and 175 MB/h for SherpaTT. This data rate for inter-robot com-munication is significantly higher compared to the field test in Utah, where the data rate was approximately 15 MB/h for SherpaTT and 11.5 MB/h for Coyote III (cf. Section6.1) This in-creased data volume is the result of the continuous exchange of footprint data (cf. Table6.3).

Table 6.3: Communication properties for the autonomous operation of an exemplary sample return mission (FM-V33).

Sender Data Volume Data Rate # of Msgs Msg Frequency Msg size

(in MB) (in MB/h) (in Hz) (in Bytes)

Coyote III 0.060 0.22 64 0.066 940±3.6

Coyote III Sensors 72.822 172.01 873 0.89 83416±218060

SherpaTT 0.007 0.03 7 0.0072 984±32

SherpaTT Sensors 46.673 268.38 924 0.95 50512±18816

Mission Performance Figure6.25visualises the performance of three trials and additional image impressions are provided in Figure I.3in Appendix I. Trial FM-V32 has been left out, since is was failed right after the payload pickup process. Trial FM-V31 failed to due a mis-alignment of Coyote III before initiating the placing of the sampling module. The implemented target alignment procedure could not compensate for the misalignment, since the target was not visible for Coyote III. Figure 6.25a shows the iterative forward movement of Coyote III leading to waypoints 5 to 8 - the approach is aborted after exceeding the permitted number of iteration.

The successful performance of a fully autonomous sample return mission sequence for FM-V33 is illustrated in Figure6.25b. Despite an orientation error of Coyote III at waypoint 4, the placing of the sampling payload succeeds. As already mentioned, the duration of the alignment is significantly higher in this trial compared to the mean performance. Also visible is a spurious pose jump of Coyote III’s assumed pose during the final backup movement. Since the robot is relying on the pose information in that situation, it has no further effect, but points to an issue of the distributed SLAM.

Figure6.25cillustrates a mission sequence, which only succeeded with operator intervention.

Two pose errors of Coyote III had to be compensated for by manually realigning the robot at waypoints 4 due to a too large orientation error, and at 9 due to overshooting - out of the manip-ulator’s workspace. The overall mission sequence ran, however, continuously and was neither explicitly halted, nor interrupted. Coyote III automatically continued to approach the target after manual correction, and likewise did the search procedure of the manipulator identify the outer marker set after moving Coyote III back into the workspace.

Figure6.24:Robot-to-robotcommunicationduringafullyautonomousmissionsequence(FM-V33).